Problems in current format. 1. IO read is slower since it needs to go for multiple seeks on the file to read column blocklets. Current size of blocklet is 120000, so it needs to read multiple times from file to scan the data on that column. Alternatively we can increase the blocklet size but it suffers for filter queries as it gets big blocklet to filter. 2. Decompression is slower in current format, we are using inverted index for faster filter queries and using NumberCompressor to compress the inverted index in bit wise packing. It becomes slower so we should avoid number compressor. One alternative is to keep blocklet size with in 32000 so that inverted index can be written with short, but IO read suffers a lot. To overcome from above 2 issues we are introducing new format V3. Here each blocklet has multiple pages with size 32000, number of pages in blocklet is configurable. Since we keep the page with in short limit so no need compress the inverted index here. And maintain the max/min for each page to further prune the filter queries. Read the blocklet with pages at once and keep in offheap memory. During filter first check the max/min range and if it is valid then go for decompressing the page to filter further. Please find the attached V3 format thrift file. Thanks & Regards,
Ravi |
Please find the thrift file in below location.
https://drive.google.com/open?id=0B4TWTVbFSTnqZEdDRHRncVItQ242b1NqSTU2b2g4dkhkVDRj On 15 February 2017 at 17:14, Ravindra Pesala <[hidden email]> wrote: > Problems in current format. > 1. IO read is slower since it needs to go for multiple seeks on the file > to read column blocklets. Current size of blocklet is 120000, so it needs > to read multiple times from file to scan the data on that column. > Alternatively we can increase the blocklet size but it suffers for filter > queries as it gets big blocklet to filter. > 2. Decompression is slower in current format, we are using inverted index > for faster filter queries and using NumberCompressor to compress the > inverted index in bit wise packing. It becomes slower so we should avoid > number compressor. One alternative is to keep blocklet size with in 32000 > so that inverted index can be written with short, but IO read suffers a lot. > > To overcome from above 2 issues we are introducing new format V3. > Here each blocklet has multiple pages with size 32000, number of pages in > blocklet is configurable. Since we keep the page with in short limit so no > need compress the inverted index here. > And maintain the max/min for each page to further prune the filter queries. > Read the blocklet with pages at once and keep in offheap memory. > During filter first check the max/min range and if it is valid then go for > decompressing the page to filter further. > > Please find the attached V3 format thrift file. > > -- > Thanks & Regards, > Ravi > -- Thanks & Regards, Ravi |
+1
This will improve the IO bottleneck. Page level min max will improve the block pruning and less number of false positive blocks will improve the filter query performance. Separating uncompression of data from reader layer will improve the overall query performance. -Regards Kumar Vishal On Wed, Feb 15, 2017 at 7:50 PM, Ravindra Pesala <[hidden email]> wrote: > Please find the thrift file in below location. > https://drive.google.com/open?id=0B4TWTVbFSTnqZEdDRHRncVItQ242b > 1NqSTU2b2g4dkhkVDRj > > On 15 February 2017 at 17:14, Ravindra Pesala <[hidden email]> > wrote: > > > Problems in current format. > > 1. IO read is slower since it needs to go for multiple seeks on the file > > to read column blocklets. Current size of blocklet is 120000, so it needs > > to read multiple times from file to scan the data on that column. > > Alternatively we can increase the blocklet size but it suffers for filter > > queries as it gets big blocklet to filter. > > 2. Decompression is slower in current format, we are using inverted index > > for faster filter queries and using NumberCompressor to compress the > > inverted index in bit wise packing. It becomes slower so we should avoid > > number compressor. One alternative is to keep blocklet size with in 32000 > > so that inverted index can be written with short, but IO read suffers a > lot. > > > > To overcome from above 2 issues we are introducing new format V3. > > Here each blocklet has multiple pages with size 32000, number of pages in > > blocklet is configurable. Since we keep the page with in short limit so > no > > need compress the inverted index here. > > And maintain the max/min for each page to further prune the filter > queries. > > Read the blocklet with pages at once and keep in offheap memory. > > During filter first check the max/min range and if it is valid then go > for > > decompressing the page to filter further. > > > > Please find the attached V3 format thrift file. > > > > -- > > Thanks & Regards, > > Ravi > > > > > > -- > Thanks & Regards, > Ravi >
kumar vishal
|
Agree.
+1 Regards JB On Feb 15, 2017, 09:09, at 09:09, Kumar Vishal <[hidden email]> wrote: >+1 >This will improve the IO bottleneck. Page level min max will improve >the >block pruning and less number of false positive blocks will improve the >filter query performance. Separating uncompression of data from reader >layer will improve the overall query performance. > >-Regards >Kumar Vishal > >On Wed, Feb 15, 2017 at 7:50 PM, Ravindra Pesala ><[hidden email]> >wrote: > >> Please find the thrift file in below location. >> https://drive.google.com/open?id=0B4TWTVbFSTnqZEdDRHRncVItQ242b >> 1NqSTU2b2g4dkhkVDRj >> >> On 15 February 2017 at 17:14, Ravindra Pesala <[hidden email]> >> wrote: >> >> > Problems in current format. >> > 1. IO read is slower since it needs to go for multiple seeks on the >file >> > to read column blocklets. Current size of blocklet is 120000, so it >needs >> > to read multiple times from file to scan the data on that column. >> > Alternatively we can increase the blocklet size but it suffers for >filter >> > queries as it gets big blocklet to filter. >> > 2. Decompression is slower in current format, we are using inverted >index >> > for faster filter queries and using NumberCompressor to compress >the >> > inverted index in bit wise packing. It becomes slower so we should >avoid >> > number compressor. One alternative is to keep blocklet size with in >32000 >> > so that inverted index can be written with short, but IO read >suffers a >> lot. >> > >> > To overcome from above 2 issues we are introducing new format V3. >> > Here each blocklet has multiple pages with size 32000, number of >pages in >> > blocklet is configurable. Since we keep the page with in short >limit so >> no >> > need compress the inverted index here. >> > And maintain the max/min for each page to further prune the filter >> queries. >> > Read the blocklet with pages at once and keep in offheap memory. >> > During filter first check the max/min range and if it is valid then >go >> for >> > decompressing the page to filter further. >> > >> > Please find the attached V3 format thrift file. >> > >> > -- >> > Thanks & Regards, >> > Ravi >> > >> >> >> >> -- >> Thanks & Regards, >> Ravi >> |
Administrator
|
Hi Ravi
Thank you bringing the discussion to mailing list, i have one question: how to ensure backward-compatible after introducing the new format. Regards Liang
|
Hi Liang,
Backward compatibility is already handled in 1.0.0 version, so to read old store then it uses V1/V2 format readers to read data from old store. So backward compatibility works even though we jump to V3 format. Regards, Ravindra. On 16 February 2017 at 04:18, Liang Chen <[hidden email]> wrote: > Hi Ravi > > Thank you bringing the discussion to mailing list, i have one question: how > to ensure backward-compatible after introducing the new format. > > Regards > Liang > > Jean-Baptiste Onofré wrote > > Agree. > > > > +1 > > > > Regards > > JB > > > > On Feb 15, 2017, 09:09, at 09:09, Kumar Vishal < > > > kumarvishal1802@ > > > > wrote: > >>+1 > >>This will improve the IO bottleneck. Page level min max will improve > >>the > >>block pruning and less number of false positive blocks will improve the > >>filter query performance. Separating uncompression of data from reader > >>layer will improve the overall query performance. > >> > >>-Regards > >>Kumar Vishal > >> > >>On Wed, Feb 15, 2017 at 7:50 PM, Ravindra Pesala > >>< > > > ravi.pesala@ > > > > > >>wrote: > >> > >>> Please find the thrift file in below location. > >>> https://drive.google.com/open?id=0B4TWTVbFSTnqZEdDRHRncVItQ242b > >>> 1NqSTU2b2g4dkhkVDRj > >>> > >>> On 15 February 2017 at 17:14, Ravindra Pesala < > > > ravi.pesala@ > > > > > >>> wrote: > >>> > >>> > Problems in current format. > >>> > 1. IO read is slower since it needs to go for multiple seeks on the > >>file > >>> > to read column blocklets. Current size of blocklet is 120000, so it > >>needs > >>> > to read multiple times from file to scan the data on that column. > >>> > Alternatively we can increase the blocklet size but it suffers for > >>filter > >>> > queries as it gets big blocklet to filter. > >>> > 2. Decompression is slower in current format, we are using inverted > >>index > >>> > for faster filter queries and using NumberCompressor to compress > >>the > >>> > inverted index in bit wise packing. It becomes slower so we should > >>avoid > >>> > number compressor. One alternative is to keep blocklet size with in > >>32000 > >>> > so that inverted index can be written with short, but IO read > >>suffers a > >>> lot. > >>> > > >>> > To overcome from above 2 issues we are introducing new format V3. > >>> > Here each blocklet has multiple pages with size 32000, number of > >>pages in > >>> > blocklet is configurable. Since we keep the page with in short > >>limit so > >>> no > >>> > need compress the inverted index here. > >>> > And maintain the max/min for each page to further prune the filter > >>> queries. > >>> > Read the blocklet with pages at once and keep in offheap memory. > >>> > During filter first check the max/min range and if it is valid then > >>go > >>> for > >>> > decompressing the page to filter further. > >>> > > >>> > Please find the attached V3 format thrift file. > >>> > > >>> > -- > >>> > Thanks & Regards, > >>> > Ravi > >>> > > >>> > >>> > >>> > >>> -- > >>> Thanks & Regards, > >>> Ravi > >>> > > > > > > -- > View this message in context: http://apache-carbondata- > mailing-list-archive.1130556.n5.nabble.com/Introducing-V3- > format-tp7609p7622.html > Sent from the Apache CarbonData Mailing List archive mailing list archive > at Nabble.com. > -- Thanks & Regards, Ravi |
Administrator
|
Hi
Thanks for detail explanation. +1 for introducing new format to improve performance further. Regards Liang
|
In reply to this post by ravipesala
hi Ravindra
As you description V3 will be benefit for IO scenairo(means more filter), what's about for CPU scenario(no filter, full scan with aggregation), is there any advantage for that. Regards Bill
|
Hi Bill,
In case of non filter query (full scan query) in V3 format carbon can read more data in single IO as we can increase number of pages in blocklet, it will reduce the IO time as number of IO will be less. -Regards Kumar Vishal On Wed, Mar 1, 2017 at 5:39 PM, bill.zhou <[hidden email]> wrote: > hi Ravindra > > As you description V3 will be benefit for IO scenairo(means more filter), > what's about for CPU scenario(no filter, full scan with aggregation), is > there any advantage for that. > > Regards > Bill > > ravipesala wrote > > Problems in current format. > > 1. IO read is slower since it needs to go for multiple seeks on the file > > to > > read column blocklets. Current size of blocklet is 120000, so it needs to > > read multiple times from file to scan the data on that column. > > Alternatively we can increase the blocklet size but it suffers for filter > > queries as it gets big blocklet to filter. > > 2. Decompression is slower in current format, we are using inverted index > > for faster filter queries and using NumberCompressor to compress the > > inverted index in bit wise packing. It becomes slower so we should avoid > > number compressor. One alternative is to keep blocklet size with in 32000 > > so that inverted index can be written with short, but IO read suffers a > > lot. > > > > To overcome from above 2 issues we are introducing new format V3. > > Here each blocklet has multiple pages with size 32000, number of pages in > > blocklet is configurable. Since we keep the page with in short limit so > no > > need compress the inverted index here. > > And maintain the max/min for each page to further prune the filter > > queries. > > Read the blocklet with pages at once and keep in offheap memory. > > During filter first check the max/min range and if it is valid then go > for > > decompressing the page to filter further. > > > > Please find the attached V3 format thrift file. > > > > -- > > Thanks & Regards, > > Ravi > > > > > > -- > View this message in context: http://apache-carbondata- > mailing-list-archive.1130556.n5.nabble.com/Introducing-V3- > format-tp7609p8137.html > Sent from the Apache CarbonData Mailing List archive mailing list archive > at Nabble.com. >
kumar vishal
|
Free forum by Nabble | Edit this page |